264 J. Opt. Soc. Am. A / Vol. 30, No. 2 / February 2013 Masaoka et al.
Number of discernible object colors is a conundrum
Kenichiro Masaoka,1,2,* Roy S. Berns,2 Mark D. Fairchild,2 and Farhad Moghareh Abed2 1NHK Science & Technology Research Laboratories, Tokyo 157-8510, Japan 2Munsell Color Science Laboratory, Chester F. Carlson Center for Imaging Science, Rochester Institute of Technology, Rochester, New York 14623, USA *Corresponding author: masaoka.k‑[email protected]
Received June 15, 2012; revised December 19, 2012; accepted December 24, 2012; posted January 2, 2013 (Doc. ID 169396); published January 31, 2013 Widely varying estimates of the number of discernible object colors have been made by using various methods over the past 100 years. To clarify the source of the discrepancies in the previous, inconsistent estimates, the number of discernible object colors is estimated over a wide range of color temperatures and illuminance levels using several chromatic adaptation models, color spaces, and color difference limens. Efficient and accurate models are used to compute optimal-color solids and count the number of discernible colors. A comprehensive simulation reveals limitations in the ability of current color appearance models to estimate the number of discernible colors even if the color solid is smaller than the optimal-color solid. The estimates depend on the color appearance model, color space, and color difference limen used. The fundamental problem lies in the von Kries-type chromatic adaptation transforms, which have an unknown effect on the ranking of the number of discernible colors at different color temperatures. © 2013 Optical Society of America OCIS codes: 330.1730, 330.4060, 330.5020, 120.5240, 230.6080, 300.6170.
1. INTRODUCTION he modified his estimate to 32,820 visual sensations (660 + The number of discernible object colors is informative not 160 + 32,000) in his revised book [4]. In 1939, Boring et al. only because of scientific interest in human color vision [5] explained that there are 156 just-noticeable differences but also because of its potential use in industrial applications (JNDs) in hue, 16–23 JNDs in saturation, and 572 JNDs in such as the development of pigments and dyes, evaluation of brightness, but he somehow made an inflated estimate of the gamut coverage of displays, and design of the spectral the number of discriminably different colors, on the order power distribution of light sources. The color gamut of color of 300,000, which is comparable to his estimate of the number imaging media and its main controlling factor, the viewing of discernible tones that differ either in pitch or in loudness. conditions, are of significant practical as well as theoretical He said, “In a sense there are, speaking generally, about as importance in the reproduction of color images [1]. However, many tones as colors.” Also in 1939, Judd and Kelly [6] esti- widely varying estimates of the number of discernible colors mated that about 10 million colors are distinguishable in day- have been made over the past 100 years. This motivated us to light by the trained human eye, although they did not mention clarify the source of the variation and derive a more accurate the grounds for their estimate and other details (e.g., criteria estimate. for color differences, assumed illuminant). This figure of 10 Perceptible color is represented in a color solid on the million was also used in the book by Judd and Wyszecki basis of the trichromatic nature of human vision, where a [7]. In 1951, Halsey and Chapanis [8] stated that a normal color solid is a three-dimensional representation of a color observer can discriminate well more than 1 million different model. Perceptual color space can be reasonably represented colors under ideal, laboratory conditions of observation. in cylindrical coordinates (hue, chroma/colorfulness, and However, in 1981, Hård and Sivik [9] stated that our ability lightness/brightness) or Cartesian coordinates (two opponent- to identify a given color with some certainty is a great deal color dimensions and lightness/brightness). The number of less and would probably cover only about 10,000–20,000 discernible colors contained in a three-dimensional solid in colors. a perceptual color space has been estimated using different In addition to these results, recent estimations were made methods. Kuehni [2] has summarized the early research. using optimal colors (i.e., optimal-color solids). An optimal Unfortunately, some of these studies do not provide sufficient color is the nonfluorescent object color having the maximum detail. To make matters worse, some estimates by authorities saturation under a given light. Ostwald [10] empirically found have been used repeatedly in other journal papers and books that optimal-color reflectances are either 0 or 1 at all wave- without reference to the original estimates, which makes it lengths with at most two transitions. The first mathematical difficult to identify the authorities. Here, we investigate the proof of this, which uses the convexity of the spectrum locus, original estimates in detail, referring to Kuehni’s summary, is attributed to Schrödinger [11]. Following the calculation of and add some recent estimates, as summarized in Table 1. optimal colors by Luther [12], Nyberg [13], and Rösch [14], In 1896, Titchener [3] experimentally identified 30,850 MacAdam generated a geometric proof of the optimal-color visual sensations (700 “brightness qualities” + 150 “spectral theorem using the CIE xy chromaticity diagram [15] and cal- colors” + 30,000 “color sensations of mixed origin”). Later, culated the chromaticity coordinates of the optimal-color loci,
1084-7529/13/020264-14$15.00/0 © 2013 Optical Society of America Masaoka et al. Vol. 30, No. 2 / February 2013 / J. Opt. Soc. Am. A 265
Table 1. History of the Estimates of the Number of Discernible Colors
Year Researcher Estimate Illuminant Color Data/Color Model 1896 Titchener [3] 30,850 700 “brightness qualities” + 150 “spectral colors” + 30,000 “colors mixed with origin” 1899 Titchener [4] 32,820 660 “brightness qualities” + 160 “spectral colors” + 32,000 “colors mixed with origin” 1939 Boring et al. [5] 300,000 156 hue × 16–23 saturation × 572 brightness 1939 Judd and Kelly [6] 10,000,000 Unknown 1943 Nickerson and Newhall [17] 7,295,000 Optimal color solid/Munsell notation 1951 Halsey and Chapanis [8] 1,000,000 Unknown 1981 Hård and Sivik [9] 10,000–20,000 Unknown 1998 Pointer and Attridge [18] 2,280,000 D65 Optimal color solid/CIELAB 2007 Wen [19] 352,263 D65 Optimal color solid/CIE94 2007 Martínez-Verdú et al. [20] 2,050,000 E Optimal color solid/CIECAM02 lightness–colorfulness (JaM bM ) 2,046,000 C 2,013,000 D65 1,968,000 F7 1,753,000 A 1,735,000 F11 1,665,000 F2 2008 Morovic [22] 1,900,000 D50 Optimal color solid/CIECAM02 lightness–chroma (JaC bC ) 2012 Morovic et al. [23] 4,200,000 F11 Optimal color solid/CIECAM02 lightness–chroma (JaC bC ) 3,800,000 D50 3,500,000 A 1,700,000 D50 LUTCHI/CIECAM02 lightness–chroma (JaC bC)
called the MacAdam limits, under illuminants A and C, with and color space, in 2012, Morovic et al. [23] doubled the num- luminous reflectance Y for the third dimension [16]. ber to 3.8 million colors under illuminant D50, in addition to In 1943, Nickerson and Newhall [17] applied the Munsell other estimates of 4.2 million under illuminant F11 and 3.5 notation to the MacAdam limits and identified a total of million under illuminant A. On the other hand, Morovic et al. 5836 full chroma steps for 40 hues spaced 2.5 hue steps apart, also pointed out that the currently available models failed to and nine values spaced one value step apart. They estimated give predictions when extrapolating the psychophysical data the number of discriminable colors as 7,295,000, assuming they are based on; finally, they made an alternative, safe es- that one chroma, hue, or value step corresponds to 5, 2, and timate of at least 1.7 million colors, which was their estimate 50 JNDs, respectively. In 1998, Pointer and Attridge [18] of the color gamut volume of the LUTCHI [24] data used to counted the unit cubes within an optimal-color solid under build CIECAM02. illuminant D65 in the CIE 1976 L ;a ;b color space Unfortunately, it is unknown which estimate is the most (CIELAB) and made an estimate of about 2.28 million. In reliable. The validity of even the estimates based on the color 2007, Wen [19] sliced an optimal-color solid under illuminant appearance models [18–20,22,23] has not been verified. To D65 at each lightness unit and divided each slice into pieces complicate matters, the recent estimates were made using having the unit chroma and hue differences described in different combinations of color appearance models (CIELAB CIE94, obtaining a count of 352,263. Also in 2007, Martínez- and CIECAM02), color spaces (lightness–chroma and Verdú et al. [20] counted unit cubes packed in optimal-color lightness–colorfulness), color difference limens (CIELAB unit, solids under several standard illuminants (A, C, D65, E, F2, F7, CIE94 color difference, and CIECAM02 unit), and lighting con- F11) and high-pressure sodium lamps in the lightness and ditions (e.g., illuminants D50 and D65). It is important to clarify colorfulness space J;aM ;bM of the CIECAM02 color appear- the source of the discrepancies in previous, inconsistent ance model [21]. They assumed this to be the most uniform estimates and determine the limitations of the current color ap- color space on the basis of their observation that the calcu- pearance models in both a theoretical and practical sense. lated optimal-color solids in the CIECAM02 color space were In this paper, we provide estimates of the number of dis- relatively spherical compared to those in the CIELAB color cernible colors within the optimal-color solid for a wide range space and others. They estimated that there are 2.050 million of color temperatures and illuminance levels, using a fast, ac- distinguishable colors under illuminant E, 2.046 million under curate model for computing the optimal colors. Several chro- illuminant C, 2.013 million under illuminant D65, 1.968 million matic adaptation models, color spaces, and color difference under illuminant F7, 1.753 million under illuminant A, and few- limens are used to verify the recent estimates based on color er colors under the other light sources. They also suggested appearance models. the possibility of an alternative color-rendering index based on the number of discernible colors within an optimal-color 2. METHODS solid. In 2008, Morovic [22] calculated the convex hull volume Figure 1 shows a flowchart of the method used to compute an of optimal-color solids in the lightness and chroma space optimal color with a specified central wavelength under an J;aC;bC of CIECAM02 and estimated a total of 1.9 million illuminant having a given spectral power distribution and colors under illuminant D50. Using the same counting method adaptation illuminance level. Optimal colors were searched 266 J. Opt. Soc. Am. A / Vol. 30, No. 2 / February 2013 Masaoka et al.
Fig. 1. Flowchart for computing an optimal color. iteratively so that their lightness values were equalized to the The spectra of the daylight illuminants were calculated using given lightness values. To estimate the number of discernible the CIE equation [29]. Following the recommendation of CIE colors within an optimal-color solid, we computed the [30], spline interpolation was used for the CIE color-matching optimal-color loci at regular lightness unit intervals of L in functions and daylight illuminant spectra in order to ensure a CIELAB, J in CIECAM02, and J0 in a CIECAM02-based uni- sufficiently small wavelength step of 0.1 nm. The blackbody form color space (CAM02-UCS) [25]. Three different chro- radiation was calculated using Planck’s law at wavelength in- matic adaptation transformations (CATs) were embedded tervals of 0.1 nm. The light source spectrum S was normalized ΣN S k y¯ k 100 N in front of CIELAB with a D65 reference white point. Black- so that k 1 , where is the number of wave- body radiators at color temperatures ranging from 2000 to length steps, N 3001. 4000 K and standard daylight illuminants at CCTs ranging To avoid switching between types I and II during optimiza- from 4000 to 10,000 K at 500 K intervals were used as light tion of λcut-on and λcut-off , three copies of the color-matching sources in the simulation. Illuminance levels of 200 lx functions x¯ k , y¯ k , and z¯ k , and illuminant spectrum S k (museum standard [26]), 1000 lx (reference viewing condition were each concatenated separately, forming four sequences [27]), 10,000 lx (outside on cloudy days), and 100,000 lx denoted respectively as x¯ c, y¯ c, and z¯c, and Sc, as shown in (outside on sunny days) were considered for the reference the upper part of Fig. 3, where n is the central wavelength, white in CIECAM02. Additionally, illuminant E and illuminant and hn is the half bandwidth of the spectral reflectance series F (F1–F12) were used at an illuminance of 1571 lx to Rn l of the optimal color on the concatenated wavelength l l n − h l n h verify the results of Martínez-Verdú et al. [20]. scale . Here, cut-on n, and cut-off n: 1;l ≤ l ≤ l A. Optimal-Color Computation cut-on cut-off Rn l ; (1) Optimal colors are generated using stimuli with either 0 reflec- 0; otherwise tance (transmittance) at both ends of the visible spectrum and 1 in the middle (type I) or 1 at the ends and 0 in the middle where n is an integer between N 1 and 2N, and hn is a real number between 0 and N∕2. It is obvious that concatenation (type II). Figure 2 illustrates both types, where λcut-on and makes it possible to treat a type II band-stop reflectance as a λcut-off represent the cut-on and cut-off wavelengths, respec- tively. In our simulation, the end colors (high-pass and low- pass) are handled as type I, with the wavelength of transition (λcut-on or λcut-off ) being the end point of the visible spectrum. A modified version of Masaoka’s model [28] was used to calculate the tristimulus values of optimal colors. We used the CIE 1931 color-matching functions x¯, y¯, and z¯ at wavelengths ranging from 400 to 700 nm at 1 nm intervals.
Fig. 3. (Color online) Concatenation of three copies of color- matching functions and illuminant spectrum (top). An optimal color Fig. 2. (Color online) Two types of spectral reflectance (transmit- has spectral reflectance R l with central wavelength n and half band- tance) for optimal colors. width hn on the concatenated wavelength scale l (center and bottom). Masaoka et al. Vol. 30, No. 2 / February 2013 / J. Opt. Soc. Am. A 267 type I band-pass reflectance. Let T x l , Ty l , and T z l be value between 0 and 1 can be set, which enables the resulting continuous functions obtained by linear interpolation of loci to be smoother than those obtained by Martínez-Verdú Scx¯ c, Scy¯ c, and Scz¯ c, respectively. Figure 4 shows a diagram et al.’s method, despite the sparse wavelength intervals. illustrating trapezoidal integration of T l . The tristimulus va- The spectral reflectance including a non-zero and non-one lues of the optimal color can be efficiently calculated using value is, however, no longer that of the optimal color by trapezoidal integration: definition. The tolerances of the optimized parameters are un- Z Z known. The 1 nm wavelength intervals degrade the accuracy l cut-off of optimal-color computation under spiky-spectrum illumi- Xn R l Tx l dl Tx l dl nants. On the other hand, our method optimizes λ and lcut-on cut-on λcut-off for an optimal color on a continuous scale rather than T x n − ⌊hn⌋ − 1 · fhng T x n − ⌊hn⌋ finding discrete reflectance values. Prevention of switching · 2 − fhng · fhng∕2 between type I and type II dramatically reduced the computa-